ConTexT-Net: Multi-Representation Fusion of Contour and Texture Features for Robust White Blood Cell Classification

초록

White Blood Cell (WBC) differential counting via blood film examination is a critical diagnostic tool for rapid and accurate clinical decision-making in systemic inflammation in animals. To address the need for automated and reliable WBC classification, this study proposes ConTexT-Net, a single-modality multi representation fusion network that achieves robust WBC differentiation by integrating contour- and texture-based representations. The proposed methodology consists of two primary stages: first, contour information is extracted using Canny edge detection to outline nuclear boundaries, and texture information is captured through adaptive thresholding to highlight fine-grained intracellular patterns. Subsequently, the preprocessed representations are fused with the RGB appearance branch using late feature-level fusion to classify WBCs. Experimental validation demonstrates that ConTexT-Net significantly outperforms a DenseNet-121 baseline, achieving an accuracy gain of 1.7 percentage points (88.9% vs. 87.2%). This enhanced performance supports the model's value as a reliable decision-support tool for veterinary clinical practice. © 2025 IEEE.

키워드